How Much Does It Cost to Build an App Like ChatGPT?
You do not build the model. You rent one from a provider such as OpenAI or Anthropic through an API, and build the product around it: accounts, chat, history, files, plans and usage limits. Here is what that costs to build, and what each message costs to run.
Test version: $8,000, about 35 days: one AI workflow with sign-up and checkout. Lean web MVP: $10,000, about 40 days: adds paid plans with message limits, and file uploads. Version 1 for teams: $13,000, 45 to 50 days: adds company workspaces and an admin dashboard of usage and model cost.
The model itself is rented from a provider through an API and billed per token. In our worked example a message costs about $0.0005 on a small model and $0.01 on a mid-size one at November 2026 list prices. For the lean build, a typical agency quote runs $70,000 to $190,000 or more by our estimate (see the MVP development cost guide).
OpenAI trains its own models; you will not. You rent a model through an API (your server sends text, the provider sends text back) and build what people pay for around it: a narrow job done well, your own instructions and content, and limits that keep usage costs below revenue. Other AI products are in our AI MVP cost guide, other app types in the app cost breakdowns.
What an MVP of an App Like ChatGPT Actually Needs
The core loop: a user signs in, asks a question or uploads a document, and gets a useful answer that streams onto the screen within seconds. They come back to the same conversation later. Behind the scenes, each user pays more than their messages cost.
- Chat with streaming. The reply appears word by word as it is written, so long answers feel instant.
- Conversation history. Saved threads the user can reopen. Each new message sends the earlier ones back to the model as context, which is why long chats cost more.
- Your instructions and knowledge. A system prompt (standing instructions the user never sees) and, where needed, search over your own content. This is what makes it your product rather than a thin wrapper.
- File uploads. PDFs and documents read into the conversation.
- Plans and limits. A free tier with a message cap and paid plans with higher caps, enforced per user.
- Safety checks. Screening of what goes in and comes out. OpenAI's moderation model is free for OpenAI API users and accepts text and images.
- Logging and cost tracking. Every request stored with its token count and cost, so you can debug a bad answer and spot expensive users.
What ChatGPT has that version one does not need
From OpenAI's list of ChatGPT Plus features:
- Image creation and voice. Separate models, built and priced apart from text chat. Add them when users ask.
- Deep research and agent mode. Multi-step tasks call the model many times per request, multiplying the cost of each one.
- Memory, projects and custom GPTs. Useful once you know how people use your product; guesswork before.
- Web search. Adds a search provider, source handling and another bill.
How ChatGPT Started
OpenAI introduced ChatGPT on November 30, 2022, as a research preview. The launch post said OpenAI had "trained a model called ChatGPT which interacts in a conversational way," fine-tuned from a model in the GPT-3.5 series that finished training in early 2022, using Reinforcement Learning from Human Feedback, in which human AI trainers played both the user and the assistant.
So the first version was a chat interface on a model fine-tuned from one OpenAI already had. "During the research preview, usage of ChatGPT is free," the post said, and it listed the limits openly, starting with "ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers." Your version one can follow the same pattern: one narrow job, an honest label on what it cannot do, and real users before more features.
Price Breakdown: Three Versions, Line by Line
Every line is a fixed price from our published price list; build days are working days. The model API connection is part of the AI product line, not a separate integration. Model usage is billed by the provider, per token, on top of the build price.
1. Test version, one AI workflow: $8,000, about 35 days
| Item | What it covers | Price | Build days |
|---|---|---|---|
| Base MVP | Sign-up and login, a chat interface with saved conversations, database, Stripe checkout or a waitlist, basic admin view, landing page, deployment, 30 days of support | $5,000 | 20 |
| AI product | The AI workflow on a hosted model: the API connection, streaming replies, prompting, evaluation, guardrails and cost control | +$3,000 | 3 |
| Total | $8,000 | 23 |
Calculator estimate: $8,000 to $9,000, about 35 days. Open this build in the calculator
What you do by hand in this version. Early users buy a one-off pass through Stripe checkout or join from a waitlist. You set each user's message cap in the admin view, users paste text instead of uploading files, and you read the conversation logs every day to see what people ask and where answers go wrong. The test asks: do people come back, and would they pay more than they cost?
2. Lean web MVP: $10,000, about 40 days
| Item | What it covers | Price | Build days |
|---|---|---|---|
| Base MVP | As in the test version | $5,000 | 20 |
| AI product | As in the test version | +$3,000 | 3 |
| Subscription billing | Free and paid plans with message limits, a trial, upgrades and a billing portal | +$1,000 | 2 |
| File processing | Upload PDFs and documents and read their text into the conversation | +$1,000 | 1 |
| Total | $10,000 | 26 |
Calculator estimate: $10,000 to $11,500, about 40 days. Open this build in the calculator
This is the version to launch publicly: plans and limits protect your margin.
3. Version 1 for teams: $13,000, 45 to 50 days
| Item | What it covers | Price | Build days |
|---|---|---|---|
| Everything in the lean MVP | Version 2 above | $10,000 | 26 |
| Teams and roles | Company workspaces, invites, admin and member roles | +$1,500 | 2 |
| Analytics dashboard | Usage and model cost by user and by team, with exports | +$1,500 | 2 |
| Total | $13,000 | 30 |
Calculator estimate: $13,000 to $15,500, 45 to 50 days. Open this build in the calculator
Sell to companies with this version: buyers want to see who uses it, and what it costs, before adding seats.
Symphsys is an AI chatbot product for homeopathy consultations, built as an MVP with production UX and deployment. The client came back for ongoing product work.
Running Costs: What Every Message Costs
Providers charge per million tokens. A token is a chunk of text, often part of a word. You pay for input tokens (your instructions, the conversation so far, any file text and the new question) and, at a higher rate, for output tokens (the reply). Standard list prices, checked November 2026:
| Model | Provider | Input, per 1M tokens | Cached input, per 1M | Output, per 1M tokens |
|---|---|---|---|---|
| gpt-6-luna | OpenAI | $0.10 | $0.01 | $0.50 |
| Claude Haiku 5.5 (prompts up to 100,000 tokens) | Anthropic | $0.10 | $0.01 | $0.50 |
| gpt-6.1-terra | OpenAI | $1 | $0.05 | $5 |
| gpt-6.1-sol | OpenAI | $2 | $0.10 | $10 |
| Claude Sonnet 5.5 | Anthropic | $2 | $0.10 | $10 |
| Claude Opus 5.5 | Anthropic | $4 | $0.20 | $20 |
| gpt-6-astra | OpenAI | $10 | $1 | $50 |
For Anthropic, the cached column is the cache-hit price; writing to the cache costs extra ($2.50 per million for Claude Sonnet 5.5's 5-minute cache). Claude Haiku 5.5 prompts over 100,000 tokens cost $0.50 and $2.50. Anthropic says its newer tokenizer produces about 30% more tokens for the same text, so equal list prices do not mean equal bills.
Worked example: 1,000 paying users
Assumptions: 1,000 users each send 10 messages a day for 30 days, so 300,000 messages a month. Each message sends 3,000 input tokens (500 of instructions, 2,400 of earlier conversation, 100 of new question) and gets 400 output tokens back. Monthly tokens: 300,000 × 3,000 = 900 million input and 300,000 × 400 = 120 million output.
| Model | Input | Output | Month | Per user |
|---|---|---|---|---|
| Small: gpt-6-luna or Claude Haiku 5.5 | 900 × $0.10 = $90 | 120 × $0.50 = $60 | $150 | $0.15 |
| Mid-size: gpt-6.1-sol or Claude Sonnet 5.5 | 900 × $2 = $1,800 | 120 × $10 = $1,200 | $3,000 | $3 |
| gpt-6-astra | 900 × $10 = $9,000 | 120 × $50 = $6,000 | $15,000 | $15 |
Per message, that is 3,000 × $2 ÷ 1,000,000 + 400 × $10 ÷ 1,000,000 = $0.006 + $0.004 = $0.01 on a mid-size model, and $0.0005 on a small one.
The heavy-user problem. Price a plan at $20 a month, and one user who sends 100 messages a day (3,000 a month) costs 3,000 × $0.01 = $30 on a mid-size model: more than they pay. Hence message limits per plan in the lean MVP, and cost per user in the team version.
Caching helps. Instructions and earlier conversation repeat every turn. If 2,000 of the 3,000 input tokens are cache reads on gpt-6.1-sol, input costs 1,000 × $2 ÷ 1,000,000 + 2,000 × $0.10 ÷ 1,000,000 = $0.0022, so a message costs $0.0062 instead of $0.01: $1,860 a month instead of $3,000. Caches expire when idle, so savings depend on how fast people reply.
The rest of the bill. Stripe charges 2.9% + 30¢ per card payment plus 0.7% for Billing: on a $20 plan, $0.58 + $0.30 + $0.14 = $1.02 per subscriber, or $1,020 a month for 1,000. Vercel Pro is $20 a month per developer seat, Supabase Pro from $25, and moderation is free for OpenAI API users. Total on a mid-size model: $3,000 + $1,020 + $20 + $25 = $4,065 a month against $20,000 of revenue; on a small model, $150 + $1,020 + $45 = $1,215. The smallest model that passes your tests is your biggest cost decision.
Timeline: The Lean MVP Week by Week
The lean MVP is 26 build days, which the calculator shows as about 40 days:
- Week 1: scope and screens agreed, skeleton deployed. You supply real example questions with good answers, which become the evaluation set (real inputs with known good outputs).
- Week 2: chat with streaming, saved conversations and the model connection, then a first evaluation run comparing a small and a mid-size model.
- Week 3: file uploads, guardrails, moderation, and logging with token counts.
- Week 4: paid plans with message limits in Stripe, usage caps and the admin view.
- Week 5: a second evaluation run, a cost check against plan prices, the landing page and testing with awkward inputs.
- Week 6: launch, handover of the source code and documentation, and 30 days of support.
More on our approach: AI development.
What Makes an AI Chat Product Hard
- Costs that grow inside each conversation. Every turn resends the history, so the twentieth message costs more than the first. Cap thread length, summarize old turns and track cost per user.
- Answers you can trust. OpenAI's own launch post warned of plausible-sounding but incorrect answers. An evaluation set turns "it seems better" into a score, and the product should decline questions its sources do not cover.
- Abuse. People will try to make your assistant ignore its instructions, and uploaded files can hide instructions too (prompt injection). Moderation, per-user rate limits and strict limits on what the model can see belong in version one.
- Depending on a provider. Models are retired and prices move, so keep the model behind one interface in your code and re-run your evaluation set before any switch.
When We Are Not the Right Choice
We do not train models. If your plan needs your own foundation model, fine-tuning at scale, models on your own hardware, or research where nobody knows whether the task can be done, you need an ML research team.
If you want a general assistant for yourself or your team, subscribe to ChatGPT or a similar product. Our fixed-price MVP model fits when a hosted model can do a narrow job and the value is in the product around it. Regulated decisions such as credit, hiring, medical or legal calls need a qualified person in the loop; we build the review step, not a replacement for it. If people chatting with each other matters more, see the team chat breakdown.
Get a Fixed Price for Your ChatGPT-Style App
Bring the job your assistant should do and ten real example questions. On a free call we will fix the build price and estimate what each user will cost you per month.
Book a Free Call →Frequently Asked Questions
How much does it cost to build an app like ChatGPT?
At minidev.pro, a lean web MVP of an AI chat app like ChatGPT costs $10,000 fixed: the $5,000 base, $3,000 of AI product work, subscription billing and file uploads. A one-workflow test version is $8,000 and a team version with a usage dashboard is $13,000. Model usage is billed separately by the provider. By our estimate, agencies typically quote $70,000 to $190,000 or more for the lean scope.
How long does it take to build an app like ChatGPT?
About 40 days for the lean MVP, which is 26 working days of build time. The test version takes about 35 days and the team version 45 to 50 days. The first week goes on collecting real example questions with good answers, because that evaluation set decides which model to use and when the product is good enough to launch.
Do I need to train my own AI model to build an app like ChatGPT?
No. You rent a model from a provider such as OpenAI or Anthropic through an API and pay per token. Training a foundation model is a research project on a different scale. The value of a new product comes from the job it does, the instructions and content you add, and the product around the model: history, files, plans and limits.
How much does each message cost in an app like ChatGPT?
About $0.0005 on a small model and $0.01 on a mid-size model in our example, where each message sends 3,000 input tokens and gets 400 back. That uses November 2026 list prices: $0.10 and $0.50 per million tokens for gpt-6-luna or Claude Haiku 5.5, and $2 and $10 for gpt-6.1-sol or Claude Sonnet 5.5. Long conversations cost more per message.
Which AI model should an app like ChatGPT use?
The smallest model that passes your evaluation set. OpenAI and Anthropic list the same prices at the small and mid sizes, so compare quality and real bills: run your example questions through both and count the tokens each bills, since Anthropic says its newer tokenizer produces about 30% more tokens for the same text. Keep the model swappable in your code.
Sources
All outside pages checked November 2026. Model prices change often; check the provider's page before you budget. For more worked examples of AI running costs, see our AI MVP cost guide.
- OpenAI, Introducing ChatGPT, November 30, 2022 (research preview; conversational model fine-tuned from the GPT-3.5 series; RLHF with human AI trainers; free during the preview; listed limitations): openai.com/index/chatgpt
- OpenAI, ChatGPT pricing (Plus features: image creation, deep research and agent mode, memory, projects and custom GPTs, voice, search): chatgpt.com/pricing
- OpenAI, API pricing (gpt-6-luna, gpt-6.1-terra, gpt-6.1-sol and gpt-6-astra input, cached input and output prices): developers.openai.com/api/docs/pricing
- Anthropic, Pricing (Claude Haiku 5.5, Sonnet 5.5 and Opus 5.5 prices; cache writes and cache hits; Haiku 5.5 prompt-length tiers; tokenizer note): platform.claude.com/docs/en/about-claude/pricing
- OpenAI, omni-moderation model ("free models designed to detect harmful content"; text and image input): developers.openai.com/api/docs/models/omni-moderation-latest
- OpenAI Help Center, Is the Moderation endpoint free to use? (free for OpenAI API users; does not count toward monthly usage limits): help.openai.com/en/articles/4936833-is-the-moderation-endpoint-free-to-use
- Stripe, Pricing (2.9% + 30¢ per domestic card charge; Billing at 0.7% of billing volume): stripe.com/pricing
- Vercel, Pricing (Pro at $20 a month per developer seat): vercel.com/pricing
- Supabase, Pricing (Pro from $25 a month): supabase.com/pricing
- minidev.pro, MVP cost calculator price list: minidev.pro/mvp-cost-calculator
- minidev.pro, Symphsys case study: minidev.pro/case-studies/symphsys